NVIDIA’s AVO Took Claude Opus 5 from 30% to 100% on ARC-AGI-3 — The Harness, Not the Model, Is the Moat

NVIDIA published a result that puts a number on where AI leverage is actually moving: the same frontier model, unchanged, went from roughly a third of a hard benchmark to a perfect score by changing the system wrapped around it.

What Happened

NVIDIA published — and The New Stack reported — that its agentic system AVO (Agentic Variation Operators), built around Anthropic’s Claude Opus 5, scored 100% on the ARC-AGI-3 public benchmark: all 183 levels, across 25 environments, in 6,624 environment actions, registering a perfect 100.00 RHAE. Two things need to be said at the front of this result, not buried. First, this is NVIDIA reporting its own system’s score — it is not an independent audit by the ARC-Prize organizers, and it should be read as NVIDIA’s published claim rather than a settled external verification. Second, 100% on the public split of ARC-AGI-3 is not AGI, it is not “reasoning solved,” and it is not a general claim about intelligence — it is one benchmark, on its public portion, and public benchmarks can be gamed or overfit by systems tuned specifically to them. Both caveats matter before anything else.

With those established: Claude Opus 5 alone had scored approximately 30% on the same benchmark — the prior high for any single model. The model’s weights did not change between that result and NVIDIA’s. What changed was the architecture wrapped around it. AVO adds persistent memory across steps, a supervision layer, and a tool-use loop that lets the system inspect and edit code, execute commands, consult documentation, and validate its own output by running it. NVIDIA’s own stated takeaway is direct: “system design, not model capability alone, can unlock frontier-level long-horizon performance.” That framing is worth taking at face value — it is the conclusion NVIDIA itself chose to lead with.

One more guard is worth stating explicitly. AVO wraps Anthropic’s model — this is not NVIDIA defeating Anthropic or OpenAI. It is complementary. NVIDIA sells compute into every lab’s stack and now publishes a reference agentic architecture; Opus 5 is the substrate that makes the score possible. The harness amplified a frontier model; it did not replace one. Opus 5 still scored 30% on its own, which means the substrate mattered: a weaker model in the same harness would not obviously reach the same ceiling. The honest framing is “system design plus a frontier model” — not “the model doesn’t matter.”

The Structural Read

The result is a single data point, but it lands at a specific moment in the AI stack’s evolution — and the timing sharpens it considerably. On the same morning this result circulated, OpenAI cut the developer price of its own frontier model by more than 20% (covered here on FWMBA). Those two events were not coordinated — they rhyme thematically, they were not planned together. But put side by side, they sketch a barbell: the model layer is visibly discounting while the layer that wraps the model is visibly capturing performance. Value is moving off the weights and onto the system.

This is the commoditization pattern the AI stack has been running for three years, and it keeps climbing. Cheap models gave way to free routing layers. Free routing gave way to discounted frontier SKUs. Now the frontier SKU is the raw material and the agentic harness is the finished product. The model is the substrate; the harness is what you sell on top of it.

NVIDIA’s move into agentic architecture is legible through this lens. The company sells the compute underneath all of this — the GPUs. If the harness is where long-horizon performance is unlocked, then owning a reference harness extends NVIDIA’s position up the stack: from selling the infrastructure into shaping the layer that determines how much that infrastructure is worth per task. As we analyzed in Beyond NVIDIA’s Moat, the GPU margin story is ultimately about whether NVIDIA can maintain relevance at each layer that sits above silicon. AVO is a bid in that direction.

There is a data dimension that sits underneath the performance headline. The harness is where the agentic traces are generated — the full record of how a system plans a multi-step task, executes tool calls, catches its own errors, and self-corrects over long horizons. That data exhaust is what future training runs will be built on. Planting a reference architecture in the harness layer is not only a benchmark headline; it is a bid for the trace data that will define the next generation of long-horizon models. The flag goes in the ground at the layer where the data is born, as mapped in The Map of AI Redrawn.

Three Implications

The Bottom Line

Discount the headline number as you should — one vendor, one benchmark, one public split — and what survives is still the part that matters: the same model, weights unchanged, went from roughly a third of a hard benchmark to the whole thing, and the entire variable was the system built around it. NVIDIA just put a specific, public number on the gap between “model alone” and “model plus harness,” and that number is not small. The layer where that gap is closed is where the moat is forming, where the trace data is generated, and where NVIDIA — a company whose core business is

91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.